Liquid AI Open-Sources Antidoom: FTPO Reduces Doom Loop Rate in Reasoning Models

Antidoom targets Doom Loops—repetitive content generation until context is full. It identifies the starting token of the loop and applies FTPO retraining only at that position to break the cycle. Official results show Doom Loop rates drop from 10.2% to 1.4% on LFM2.5-2.6B and from 22.9% to 1% on Qwen3.5-4B. The full Antidoom suite (generation, detection, FTPO trainer) is open-sourced. For teams finetuning reasoning models (e.g., long-chain reasoning, multi-step agents), Antidoom directly solves loop pitfalls that previously required manual intervention. With localized FTPO retraining, model stability improves at low cost, especially beneficial in resource-constrained deployments like edge devices.